Hi, I am currently endeavoring to train the joint SDF-VAE to acquire SDF modulations. Successful reconstruction is observed when processing data in the object normalization space (as exemplified by ShapeNet). However, when I rotate and translate the object for training, but also normalize the object in the unit cube space. Despite a seemingly well-performing loss, the model struggles to faithfully reconstruct the object. Presented below are the reconstructed results; could this discrepancy be indicative of a limitation in the model?
Hi, I am currently endeavoring to train the joint SDF-VAE to acquire SDF modulations. Successful reconstruction is observed when processing data in the object normalization space (as exemplified by ShapeNet). However, when I rotate and translate the object for training, but also normalize the object in the unit cube space. Despite a seemingly well-performing loss, the model struggles to faithfully reconstruct the object. Presented below are the reconstructed results; could this discrepancy be indicative of a limitation in the model?